Light-based chip claims 100× faster, more efficient performance than Nvidia A100

Researchers from Shanghai Jiao Tong University and Tsinghua University have unveiled "LightGen," a breakthrough all-optical computing chip designed specifically to handle the immense demands of generative artificial intelligence. The chip — detailed in the journal Science — represents a significant shift from electronic transistors to photonic neurons, offering a potential solution to the massive energy bottlenecks currently facing the AI industry.

While previous optical processors contained a few thousand neurons and were largely limited to simpler tasks like image classification, LightGen utilizes advanced 3D packaging to integrate over two million artificial neurons into a single quarter-square-inch device. This massive scale allows the chip to execute complex generative tasks, such as high-definition video generation and 3D modeling, which were previously the exclusive domain of high-end electronic GPUs.

A core innovation of the design is the "optical latent space." By using ultra-thin metasurfaces and optical fiber arrays, the chip can compress and process high-dimensional data entirely through light. This allows the system to work with full-resolution images without breaking them into patches, preserving vital statistical data and dramatically increasing throughput. The researchers reported that the chip's performance is over 100 times faster than a leading Nvidia A100 GPU.

In laboratory tests, LightGen successfully performed high-resolution semantic image generation and 3D manipulation with quality comparable to leading electronic neural networks. While the technology currently relies on external laser setups and specialized manufacturing processes, it provides a promising roadmap for the future of high-speed, sustainable, intelligent computing.

LightGen opens a new path for advancing generative AI with higher speed and efficiency, providing a fresh direction for research into high-speed, energy-efficient generative intelligent computing. — Yitong Chen, lead author of the paper.